Applying deep learning and eXplainable AI (XAI) techniques to medical applications

Tan, Keng Ze (2026) Applying deep learning and eXplainable AI (XAI) techniques to medical applications. Final Year Project, UTAR.

PDF Download (1700Kb)

Abstract

Incorporation of AI in medical diagnosis continues to evoke apprehensions due to the black-box nature of deep learning models. Clinicians get a prediction result without comprehending which regions of the image influenced the model's decision. The lack of model prediction interpretability constrains trust and practical implementation in clinical environments. This project aims to create a breast ultrasound lesion detection system based on Faster R-CNN with ResNet50-FPN utilizing the BUSI dataset. Contrary to pure image classification, the system concentrates on object detection by predicting the locations of bounding-boxes and class labels for benign, malignant, and normal cases. Segmentation masks from the dataset are transformed into bounding boxes to facilitate detection training. Grad-CAM visualisation is employed to assess the alignment of model attention with pertinent lesion areas. Furthermore, various masked filter pruning strategies are adapted into Faster R-CNN to evaluate the filter-importance, including random pruning, activation-only pruning, Taylor-only pruning, hybrid activation-Taylor pruning, and loss-guided Grad-CAM-style relevance pruning. Experimental results show that the baseline Faster R-CNN model attained satisfactory detection performance, with mAP@0.50 ranging from 0.70 to 0.74 and mAP@0.75 between 0.53 and 0.64 across multiple runs. Taylor-only and hybrid activation-Taylor pruning demonstrated competitive performance in specific contexts. However, the pruning outcomes should be regarded as a filter-importance ablation rather than significant model compression, as the pruning was masked and implemented solely on one designated layer. The parameter reduction was minimal, ranging from 0.25% to 1.78%, and the enhancement in inference speed was constrained. Grad-CAM visualisation offered qualitative explanatory support, although the heatmaps were not consistently well-aligned with lesion areas. A cloud-based deployment prototype was developed to showcase model accessibility via a web platform. This project introduces a development-oriented breast ultrasound detection pipeline that integrates localisation, explainability, pruning analysis, and deployment viability.

Actions (login required)